Is there a way that I can give more importance to the new tokens that I ve added to the vocabulary, besides creating my own trainer and using my own loss function?
2 Likes
Have you seen logit_bias or LogitsProcessor?
1 Like